Developing an OpenAI Gym-compatible framework and simulation environment for testing Deep Reinforcement Learning agents solving the Ambulance Location Problem

TL;DR

This paper introduces an OpenAI Gym-compatible simulation environment for testing Deep Reinforcement Learning agents on the Ambulance Location Problem, demonstrating potential improvements in response times over random allocation methods.

Contribution

It presents a novel, customizable simulation framework for applying Deep RL to ambulance dispatch optimization, facilitating future research and development.

Findings

Deep RL agents reduced response times compared to random allocation.

Bagging Noisy Dueling Deep Q networks showed the most consistent performance.

Agents' performance declined if trained excessively, highlighting the need for optimal training duration.

Abstract

Background and motivation: Deep Reinforcement Learning (Deep RL) is a rapidly developing field. Historically most application has been made to games (such as chess, Atari games, and go). Deep RL is now reaching the stage where it may offer value in real world problems, including optimisation of healthcare systems. One such problem is where to locate ambulances between calls in order to minimise time from emergency call to ambulance on-scene. This is known as the Ambulance Location problem. Aim: To develop an OpenAI Gym-compatible framework and simulation environment for testing Deep RL agents. Methods: A custom ambulance dispatch simulation environment was developed using OpenAI Gym and SimPy. Deep RL agents were built using PyTorch. The environment is a simplification of the real world, but allows control over the number of clusters of incident locations, number of possible dispatch locations, number of hospitals, and creating incidents that occur at different locations throughout each day. Results: A range of Deep RL agents based on Deep Q networks were tested in this custom environment. All reduced time to respond to emergency calls compared with random allocation to dispatch points. Bagging Noisy Duelling Deep Q networks gave the most consistence performance. All methods had a tendency to lose performance if trained for too long, and so agents were saved at their optimal performance (and tested on independent simulation runs). Conclusions: Deep RL agents, developed using simulated environments, have the potential to offer a novel approach to optimise the Ambulance Location problem. Creating open simulation environments should allow more rapid progress in this field.